Meituan unwraps LongCat-2.5, a 1.6T-parameter model for AI that does the chores

Meituan pulled the wrapper off LongCat-2.5 on November 2, 2026, and it's available now through an API. The model carries 1.6 trillion total parameters, though only about 48 billion are active per token, which keeps the compute bill in check. Meituan built it for long-horizon agentic tasks, the kind of multi-step jobs where the AI keeps working on its own instead of waiting for the next question.

The practical reach is wide. LongCat-2.5 runs a native context window of one million tokens and can generate outputs up to 128,000 tokens. Meituan says that lets it operate across terminals, browsers, graphical interfaces, spreadsheets and design tools — agentic coding and visual work, not chit-chat, is the pitch.

Two upgrades separate it from LongCat-2.0, which landed in June 2026. First, vision: the model chops images into multi-level discrete tokens and reasons about them in the same stream as text, rather than treating a picture as a file bolted on the side. Second, training: Meituan leaned on long-horizon reinforcement learning built around failure trajectories and synthetic data, and squeezed inference with W4A4 quantization that drops weights and activations to 4-bit precision, which the company says delivered substantial throughput gains.

The full launch followed a September 25, 2026 preview that showed off image understanding but shipped without performance benchmarks. November fixed that. Meituan's own numbers put LongCat-2.5 competitive with Claude Opus variants on several benchmarking tasks — a claim that, for now, rests on Meituan's reporting.

One spec stands out from the rest. LongCat-2.5 was trained on domestically produced Chinese AI accelerators, fitting Meituan's push for technological independence as US export restrictions on advanced AI tech keep tightening. For a company famous for fast dumpling delivery, the subtext is clear: it wants to move just as fast on its own silicon.

Disclosure: Crypto Briefing noted the article was edited by Diego Almada Lopez.